Analog Vlsi Synapse Matrix With Enhanced Stochastic Computations
Michel Verleysen, P. Jespers · Lecture notes in computer science · 1991
Most of the applications of neural networks require large arrays of synapses and neurons. Generally, neural networks are used in problems when classical algorithms are so complex that it is difficult or impossible to use them efficiently in real-time applications. When implementing the network on silicium, digital techniques can be used, with a great precision in the computations but with large cells which are to be multiplexed, or analog ones, which are faster and smaller, but which present drawbacks in their precision and cascadability because of the mismatching between components. This paper presents an architecture which offers a good compromise between analog and digital layouts, by using stochastic computations in small synapses with good precision.